Every model has two versions: a base model (raw pre-training, mostly useful for research) and an instruct model (fine-tuned via RLHF or similar to follow instructions, refuse harmful requests, and behave conversationally). When you call the OpenAI API, use Claude, or run Llama locally, you're almost always using the instruct version.
The distinction matters when you're fine-tuning: fine-tuning on top of an instruct model preserves the safety and instruction-following behavior. Fine-tuning on a base model gives you more flexibility but loses those guardrails. Most production teams fine-tune instruct models.
Also relevant: instruct models are often slightly worse than base models on pure benchmarks because alignment can reduce raw capability. This is the "alignment tax" — usually worth paying, but worth knowing exists.
Bring this to your business
Knowing the term is one thing. Shipping it is another.
We do two-week AI Sprints — one term, one workflow, into production by Day 10.